owned this note
owned this note
Published
Linked with GitHub
# Default index
## TL;DR
The proposal is to have an option `mode.no_default_index`, in which:
- by default, DataFrames will be created without an Index
- users will never end up with an index unless they ask for one (https://github.com/pandas-dev/pandas/issues/49069)
```python
In [3]: with pd.option_context('mode.no_default_index', True):
...: df = pd.DataFrame({'a': [1, 2, 3], 'b': [4, 5, 6]})
...:
In [4]: df
Out[4]:
a b
1 4
2 5
3 6
```
## Why?
The Index can be a source of confusion and frustration for pandas users. For example, let's consider the inputs
```python
In [37]: ser1 = df.groupby('sender')['amount'].sum()
In [38]: ser2 = df.groupby('receiver')['amount'].sum()
In [39]: ser1
Out[39]:
sender
1 10
2 15
3 20
5 25
Name: amount, dtype: int64
In [40]: ser2
Out[40]:
receiver
1 10
2 15
3 20
4 25
Name: amount, dtype: int64
```
. Then:
- it can be unexpected that summing `Series` with the same length (but different indices) produces `NaN`s in the result ( https://stackoverflow.com/q/66094702/4451315):
```python
In [41]: ser1 + ser2
Out[41]:
1 20.0
2 30.0
3 40.0
4 NaN
5 NaN
Name: amount, dtype: float64
```
- concatenation, even with `ignore_index=True`, still aligns on the index (https://github.com/pandas-dev/pandas/issues/25349):
```python
In [42]: pd.concat([ser1, ser2], axis=1, ignore_index=True)
Out[42]:
0 1
1 10.0 10.0
2 15.0 15.0
3 20.0 20.0
5 25.0 NaN
4 NaN 25.0
```
- it can be frustrating to have to repeatedly call `.reset_index()` (https://twitter.com/chowthedog/status/1559946277315641345):
```python
In [45]: df.value_counts(['sender', 'receiver']).reset_index().rename(columns={0: 'count'})
Out[45]:
sender receiver count
0 1 1 1
1 2 2 1
2 3 3 1
3 5 4 1
```
With this mode enabled, two major changes would happen:
- by default, a DataFrame would be created without an Index;
- nobody would get an index unless they ask for one. This would involve changing the default `as_index` option in `groupby`, and allowing for `value_counts` to not set an index.
With this option enabled, users who don't want to worry about indices could safely ignore them.
## How?
### NoIndex DataFrame
A DataFrame without an index would have an index which would behave like a RangeIndex, except for the following differences:
- `name` could only be `None`;
- `start` could only be `0`, `step` `1`;
- when appending an extra element, the new `Index` should still be `NoIndex`;
- when slicing, one should still get a `NoIndex`;
- when printing a DataFrame, the row labels should be hidden.
- two no-index objects shouldn't be aligned. Either they're the same length, or pandas raises;
- aligning a no-index object with one which has an index will raise, always;
- columns would not be allowed to be `NoIndex` (so `transpose` would need some adjustments);
- `insert` and `delete` should raise. In particular, `.drop` with `axis=0` would aways raise;
- arithmetic operations should probably all raise;
### Don't give people an index unless they ask for one
Some pandas methods create an Index by default. This can sometimes be opted out of (e.g. with `as_index=False` in `.groupby`), but other times there is no choice but to call `reset_index` after the operation (e.g. with `.pivot_table` and `.value_counts`).
A couple of solutions come to mind:
- add `as_index` options to these methods, whose default could be `False` under this option;
- in this option, the behaviour of these methods would change and no index would be introduced.
The second would keep API size down, whilst the first one would give the most flexibility to users. I'd be more inclined towards the former.
### How to ask for an index?
It should be fine to do `df.reset_index().set_index('index')`, no need to add a new method.
## Downstream libraries
### seaborn
Seaborn makes extensive use of label-based indexing, and so NoIndex DataFrames would break it:
```
In [1]: df = pd.DataFrame({'a': [1, 1, 2], 'b': [1, 3, 4]})
In [2]: import seaborn as sns
In [3]: sns.lineplot(df)
NotImplementedError: Can't reindex a DataFrame without an index. First, give it an index.
```
Even if `df` had an Index, `seaborn.lineplot` would still error because internally it creates new DataFrames (which now wouldn't have an index) and then it would call things that wouldn't work on them, such as `data.loc[[]]`.
This would need some working out.
## Why not have `.index` be None, rather than a NoIndex?
`.index` methods are quite common to call, e.g.
https://github.com/pandas-dev/pandas/blob/dbb2adc1f353d9b0835901c274cbe0d2f5a5664f/pandas/core/series.py#L877
in
```python
ser = Series([1,2,3])
breakpoint()
ser.loc[ser>1]
```
## Roadmap - how to make this change?
In pandas 2.x.0, introduce the `mode.no_default_index` option. It's unlikely that this could ever be made the default, but it could be made the default in a separate namespace (which would try to be compliant with the DataFrame standards API).
## Resources
pandas issue: https://github.com/pandas-dev/pandas/issues/48880